NVIDIA's Spark local hardware lineup was updated across the board at GTC Berlin on October 21, 2026. In its official blog post that day, NVIDIA walked through RTX Spark, DGX Spark and DGX Station, with one core change: local agents are no longer a single-machine demo. DGX Spark now clusters up to eight units for 1TB of unified memory — enough to run frontier-class models at the desk.
On a single unit, the DGX Spark 64GB systems shipping from partners starting October 23 can run models such as Nemotron 3.5 Bolt 74B 24GB and Qwen3.8 27B. An eight-unit cluster, linked by built-in ConnectX-7 and NVIDIA Sync, can take on larger coding models like Qwen 3.8 Flash Next and DeepSeek v4 Flash. RTX Spark targets personal agents on Windows PCs with up to 128GB of unified memory and 1 petaflop of AI compute, and the new Underdog app already runs fully locally so data never leaves the PC.
For enterprise teams, DGX Station offers 748GB of coherent memory and up to 20 petaflops, serving 30-plus users, with seven new enterprise management integrations this month and a Windows version expected in December. The local coding model Recursive Flint, released the same day, tops 220 tokens per second on a 64GB RTX Spark, with weights open under Apache 2.0.
The point of this lineup is to unbundle privacy, latency and cost from the cloud bill: teams whose code, documents and customer data must stay on-premises can validate at small scale first, then decide which workloads are worth bringing back local. The cost is equally clear — cluster networking, model quantization and operations are on you. Local does not mean maintenance-free.
